detour-ninja/detours-eavesdrop

Detours Eavesdrop is a local first TypeScript workbench that takes a YouTube video URL and converts the video into structured engineering knowledge. It extracts the content, transcripts, coding skills, step by step guides, rules, prompts, hooks, agent patterns, and code snippets, and lets you push the results directly into Claude Code, Codex, or a brand new GitHub repository as a starter workspace. Everything runs on your own machine. Analyses and history are stored in a local SQLite database, so you never pay twice to analyse a video you have already processed. There is no server side storage and no built in AI keys, you supply your own credentials, bring your own key, for any of five supported providers: OpenAI, Anthropic, OpenRouter, Eliza Cloud, or ChatGPT via Codex OAuth. A live model dropdown fetches the current model list directly from each provider and caches it for 24 hours, so you always see up to date model names. You can use it two ways: as a web UI served locally at port 3000, with keyboard shortcuts and a history sidebar, or as an interactive command line tool that auto detects whichever provider credentials are already configured on your machine. The same analysis engine is also exposed as a REST endpoint and a Model Context Protocol server for programmatic or agent driven use. The stack is Bun for the runtime and HTTP server, Vite and React for the browser UI, bun:sqlite for the local database, and Tailwind for styling. The full README is longer than what was shown.